基于不同时相增强CT的影像组学预测≤3 cm非小细胞肺癌脏层胸膜侵犯:一项双中心研究  

Radiomics-based prediction of visceral pleural invasion in non-small cell lung cancer with enhanced CT at different time phases≤3 cm:A two-center study

作　　者：洪晨晨 李建斌[2] 陈炜越 刘柏君 冯烨 郑毅 陈洁[4] 纪建松[1,4] 涂建飞[1,4] HONG Chenchen;LI Jianbin;CHEN Weiyue;LIU Bojun;FENG Ye;ZHENG Yi;CHEN Jie;JI Jiansong;TU Jianfei(Zhejiang Key Laboratory of Imaging and Interventional Medicine,the Fifth Affiliated Hospital of Wenzhou Medical University,Lishui Central Hospital,Lishui 323000,China;Department of Radiology,the Affiliated People’s Hospital of Ningbo University,Ningbo 315211,China;Suzhou Institute of Biomedical Engineering and Technology,Chinese Academy of Sciences,Suzhou 215163,China;Department of Intervention,the Fifth Affiliated Hospital of Wenzhou Medical University,Lishui Central Hospital,Lishui 323000,China)

机构地区：[1]温州医科大学附属第五医院(丽水市中心医院)全省影像与介入医学重点实验室,浙江丽水323000 [2]宁波大学附属人民医院放射科,浙江宁波315211 [3]中国科学院苏州生物医学工程技术研究所,江苏苏州215163 [4]温州医科大学附属第五医院(丽水市中心医院)介入科,浙江丽水323000

出　　处：《温州医科大学学报》2024年第12期987-995,共9页Journal of Wenzhou Medical University

基　　金：国家卫生健康委员会科研基金项目(WKJ-ZJ-2452)。

摘　　要：目的:探讨基于不同时相增强CT影像组学模型在术前预测≤3 cm非小细胞肺癌(NSCLC)脏层胸膜侵犯(VPI)的价值。方法:选取2019年1月至2023年9月在温州医科大学附属第五医院及宁波大学附属人民医院经手术病理证实为NSCLC并接受增强CT检查的患者325例。根据病理结果,将患者分为VPI阳性组和VPI阴性组。将温州医科大学附属第五医院数据集经随机分层抽样法按7:3的比例分为训练集(163例,VPI阴性/VPI阳性为104例/59例)、内部测试集(70例,VPI阴性/VPI阳性为44例/26例),宁波大学附属人民医院病例作为外部测试集(92例,VPI阴性/VPI阳性为62例/30例)。通过人工智能辅助诊断建模分析软件分别基于平扫期(NP)、动脉期(AP)、静脉期(VP)及NP+VP+AP提取图像中的影像组学特征,依次采用方差分析选择法、单变量特征选择法以及最小绝对收缩和选择算子进行降维,筛选出具有鉴别意义的组学特征,并构建了各时相的单独模型及多时相模型。单因素比较差异有统计学意义的临床资料和CT形态学特征,结合最终相关性最高的影像组学特征构建融合模型。最后通过ROC曲线评估不同模型的预测效能。结果:从NP、AP、VP及NP+VP+AP图像中筛选得到8、9、8和13个最优影像组学特征,分别构建NP、AP、VP单独时相模型和NP+AP+VP多时相模型。ROC曲线结果显示,NP+AP+VP模型的AUC值为0.891,优于任一单独时相模型。VPI阴性组与VPI阳性组间肿瘤最大径和肿瘤密度差异均有统计学意义(均P<0.05)。融合模型由2种CT形态学特征与13个最优的影像组学特征共同构建,在训练集和内/外部测试集中展示出最高的预测效能,其AUC值分别为0.906、0.85和0.884,优于其他模型。结论:基于多时相增强CT图像的影像组学融合模型对≤3 cm的NSCLC患者VPI状态具有较好的预测价值,可为临床的治疗与决策提供重要参考。Objective:To explore predictive models constructed from multi-phase contrast-enhanced CTbased radiomics for preoperatively predicting visceral pleural invasion(VPI)in≤3 cm non-small cell lung cancer(NSCLC).Methods:Retrospectively,325 patients diagnosed with NSCLC confirmed by surgical pathology and undergoing enhanced CT examinations at the Affiliated Fifth Hospital of Wenzhou Medical University and the Affiliated People’s Hospital of Ningbo University between January 2019 and September 2023 were selected.Based on pathological findings,patients were categorized into VPI-positive and VPI-negative groups.The dataset from our hospital was randomly divided using a stratified sampling method into a training set(163 cases,with 104 VPI-negative and 59 VPI-positive cases),an internal test set(70 cases,with 44 VPI-negative and 26 VPIpositive cases),while the cases from the Affiliated People’s Hospital of Ningbo University served as an external test set(92 cases,with 62 VPI-negative and 30 VPI-positive cases).Artificial intelligence assisted diagnostic modeling and analysis software were used to extract radiomic features from images based on plain scan(NP),arterial phase(AP),venous phase(VP),and NP+VP+AP phases.The variance analysis selection method,univariate feature selection method,and minimum absolute shrinkage and selection operator were used in sequence for dimensionality reduction to select omics features with discriminative significance.Separate and multitemporal models were constructed for each phase.Univariate comparisons of clinically significant data and CT morphological features were made,and a fusion model was built incorporating the radiomic features with the highest relevance.Finally,the predictive performance of each model was evaluated using the ROC curve.Results:From the NP,AP,VP,and combined NP+VP+AP images,8,9,8,and 13 optimal radiomic features were selected,respectively.Separate phase models for NP,AP,VP and a comprehensive model combining NP+AP+VP were constructed.The ROC results showed that th

关 键 词：早期非小细胞肺癌 脏层胸膜侵犯 影像组学 体层摄影术 X线计算机 

分 类 号：R445.3[医药卫生—影像医学与核医学]